OpenAI is freezing its latest frontier run

Welcome back. As models become more capable, the risks associated with them also go up. OpenAI just hit a massive milestone (or a major roadblock), forcing them to freeze training for two full weeks. They're calling it a safety play to avoid incidents like the Hugging Face breach, but this could also be linked to their new model's exceptional abilities. Read the details below.

Also: A dev’s weekend project that blew up on the internet, how to hire and keep talented devs, and how to build an AI-native PR review.

OpenAI is pausing model development due to cybersecurity concerns: The ChatGPT maker paused two weeks of model training to tighten security, and its largest planned reinforcement learning (RL) run is still on hold. The move came as a result of the Hugging Face incident that occurred last month during model testing. This suggests the upcoming Astra model likely crossed a critical cyber capability threshold. OpenAI is freezing many Astra and cybersecurity workloads until they pass tougher checks while adding earlier monitoring and stronger safeguards across the entire pipeline. More details here.

Warp ships plug-and-play infrastructure for software factories: This AI coding startup just launched a system that makes building software factories effortless. Developers can now deploy fleets of coding agents across their entire SDLC. You can use your preferred model and harness, whether it's Codex or Claude Code, and it integrates directly with Linear, Jira, Slack, and Teams. You can try it here.

Frontier labs compete to slash costs for developers: OpenAI just slashed prices on its flagship model across different gateways. GPT-5.6 Sol is now 70% off on Cognition’s Devin through October 3 and reportedly 50% off on OpenRouter and Vercel’s AI Gateway. And Anthropic isn’t sitting this one out either, they’re boosting Claude Code weekly limits by 50% through the end of August. Either way, it’s a win for devs looking to build heavy projects in the coming months.

Soniox Text-to-Speech lets you create extraordinary voice experiences in 60+ languages with expressive control, exceptional precision, instant voice cloning, and ultra-low-latency streaming.

Direct whispers, laughter, hesitation, excitement, and emotion, or precisely speak phone numbers, email addresses, technical terminology, and multilingual text.

Built for voice agents, global products, customer support, media, and any application where speech needs to sound human and get every detail right.

$0.70 per generated hour.

Try it now

False Confidence. Coding agents sound just as confident when they're wrong as when they're right. If you take a suggestion at face value, you risk shipping a mistake. The gap between an answer that looks good and one that actually works is exactly what senior developer Kent Dodds explores in his podcast. In his workflow, the agent digs up platform notes, runs tests, and offers suggestions while the engineer makes the final judgment call. 

The tip that would've cost 13x. For instance, an agent suggested storing large files in Cloudflare Durable Objects and dismissed the cost as pocket change. Dodds had a gut feeling it would be pricier, so he pushed back. He was right, his math showed it would cost about 13 times more than using R2, Cloudflare’s dedicated storage for large files. That catch wasn't luck. It was found because, as part of his workflow, he pushes every agent suggestion through a few simple questions before trusting it:

Additionally, he doesn't just rely on gut feeling. He leans on the agents pulling the hard numbers:

The real value. Knowing which suggestions to trust, toss, or question is becoming a core engineering skill. As Dodds says, the engineers who do this best are often the most valuable people on the team. 

PS: For the skill underneath it all, this podcast works through product judgment with engineers and PMs.

Take the Tiger Hunt: 5 hands-on challenges in Tiger Cloud. Load data, use the CLI or MCP, build a hypertable, create a continuous aggregate, and join a workshop.

Finish the hunt and earn a Tiger tee, hat + stickers.

Ready to hunt?

How to build an AI native PR review system: You’ll learn how to build a production-ready, multi-agent AI PR review system. This tutorial also covers fault-tolerant architecture, database modeling, and agentic orchestration to automate reviews with senior-level engineering judgment. 

Social-fetch: This skill lets AI agents read social posts from X, LinkedIn, Instagram, TikTok, Reddit, and HN, getting around login walls with APIs, browser automation, or Wayback.

Hands on guide to loop engineering (by a Google engineer): Agents often fail by repeating mistakes or lacking oversight. Loop engineering fixes this with self-correcting workflows. This guide covers building verification steps, automating PR triage with Claude Code, and setting clear stopping conditions for reliable agent autonomy. 

You know the drill. Claude is halfway through a big refactor, the usage limit kicks in, and you come back hours later to a dead session.

Anthropic's ClaudeDevs account just shared the fix, an auto-continue checkbox in Claude Code desktop.

In the demo, a payments migration finished edits and tests with zero handholding.

P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.

Check out how a Google engineer found six ways developers quietly burn tokens in Claude Code, then fixed them with a single slash command.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 300K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team